IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO.
2, APRIL 2008
3-D Snake Robot Motion: Nonsmooth Modeling,
Simulations, and Experiments
Aksel A. Transeth, Member, IEEE, Remco I. Leine, Christoph Glocker,
and Kristin Y. Pettersen, Senior Member, IEEE
Abstract—A nonsmooth (hybrid) 3-D mathematical model of
a snake robot (without wheels) is developed and experimentally
validated in this paper. The model is based on the framework of
nonsmooth dynamics and convex analysis that allows us to easily
and systematically incorporate unilateral contact forces (i.e., be-
tween the snake robot and the ground surface) and friction forces
based on Coulomb’s law of dry friction. Conventional numerical
solvers cannot be employed directly due to set-valued force laws
and possible instantaneous velocity changes. Therefore, we show
how to implement the model for numerical treatment with a nu-
merical integrator called the time-stepping method. This method
helps to avoid explicit changes between equations during simula-
tion even though the system is hybrid. Simulation results for the
serpentine motion pattern lateral undulation and sidewinding are
presented. In addition, experiments are performed with the snake
robot “Aiko” for locomotion by lateral undulation and sidewinding,
both with isotropic friction. For these cases, back-to-back compar-
isons between numerical results and experimental results are given.
Index Terms—3-D snake robot, kinematics, nonsmooth dynam-
ics, time-stepping method.
I. INTRODUCTION
HEELED mechanisms constitute the backbone of most
W ground-based means of transportation. Unfortunately,
rough terrain makes it hard, if not impossible, for such mecha-
nisms to move. To be able to move in various terrains, such as
going through narrow passages and climb on rough ground, the
high-mobility property of snakes is recreated in robots that look
and move like snakes.
Snake robots most often have a high number of DOFs, and
they are able to move forward without using active wheels or
legs. Due to the high number of DOF, it can be quite expensive
and time-consuming to build and maintain a snake robot. This
motivates the development of accurate mathematical models of
snake robots. Such models can be used for synthesis and testing
of various serpentine motion patterns intended for serpentine
locomotion.
Manuscript received July 19, 2007; revised November 29, 2007. This paper
was recommended for publication by Associate Editor S. Ma and Editor F. Park
upon evaluation of the reviewers’ comments.
A. A. Transeth is with the Department of Applied Cybernetics, SINTEF Infor-
mation and Communication Technology (ICT), NO-7465 Trondheim, Norway
(e-mail: aksel.transeth@sintef.no).
R. I. Leine and Ch. Glocker are with the Institut für Mechanische Systeme
(IMES) Centre of Mechanics, Eidgenössische Technische Hochschule (ETH)
Zürich, CH-8092 Zurich, Switzerland (e-mail: remco.leine@imes.mavt.ethz.ch;
christoph.glocker@imes.mavt.ethz.ch).
K. Y. Pettersen is with the Department of Engineering Cybernetics,
Norwegian University of Science and Technology, NO-7491 Trondheim,
Norway (e-mail: kristin.y.pettersen@itk.ntnu.no).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TRO.2008.917003
1552-3098/$25.00 © 2008 IEEE
361
The first working snake robot was built in 1972 [1]. This
robot was limited to planar motion, but snake robots capable
of 3-D motion have appeared more recently [2]–[6]. Together
with the robots, mathematical models of both the kinematics
and the dynamics of snake robots have also been developed.
Purely kinematic 3-D models have been presented in [6]–[8],
where frictional contact between the snake robot and the ground
is not included in the model. Hence, contact between the snake
robot and the ground surface is either modeled with frictionless
passive wheels, or the parts of the snake robot that touches the
ground are defined as anchored to the ground [9]. A model
of the dynamics of motion is needed to describe the friction
forces a snake robot without wheels is subjected to when moving
over a surface. Most mathematical models that describe the
dynamics of snake robot motion have been limited to planar
(2-D) motion [10]–[13], and 3-D mathematical models of snake
robots have only recently been developed [3], [7]. 3-D models
facilitate testing and development of 3-D serpentine motion
patterns such as sinus lifting and sidewinding. A description of
these motion patterns is found in, e.g., [4]. A physics engine
called the open dynamics engine (ODE) has been employed
to simulate a 15-link snake robot instead of deriving explicit
expressions for its dynamics in [14]. Such software makes it
easy to change the geometry of a snake robot if needed, and the
time needed to prepare a working model is relatively short [15].
On a flat surface, the ability of a snake robot to move forward
is dependent on the friction between the ground surface and the
body of the snake robot. Hence, unilateral contact forces and
friction forces are important parts of the mathematical model of
a snake robot. The friction forces have usually been based on a
Coulomb or viscous-like friction model [11], [12], and Coulomb
friction has most often been modeled using a sign function [12],
[16]. The contact between a snake robot and the ground surface
can sometimes be approximated by a no-sideways-slip con-
straint for snake robot with wheels [17], [18]. However, such
an approximation is not valid for wheel-less snake robots. The
unilateral contact forces have been modeled as a mass-spring-
damper system in [3] (i.e., compliant contact), and each link has
only a single and fixed contact point with the ground surface.
When running simulations, direct implementations or approxi-
mations of the sign function can lead to an erroneous description
of sticking contacts or very stiff differential equations. Also, a
mass-spring-damper model introduces a very stiff spring that
leads to stiff differential equations. In addition, it is not clear
how to determine the dissipation parameters of the contact un-
ambiguously when using a compliant model [19]. The ODE
implements a form of rigid body contact (i.e., not compliant
contact). However, the implementation of this engine trades off
362
Fig. 1. NTNU/SINTEF snake robot “Aiko.”
simulation accuracy in order to increase simulation speed and
stability [15], [20].
In this paper, we develop a nonsmooth (hybrid) 3-D mathe-
matical model of a snake robot with cylindrical links without
wheels. Set-valued force laws for the constitutive description of
unilateral contact forces and friction forces in a 3-D setting are
described in the framework of nonsmooth dynamics and convex
analysis [19], [21], [22]. Moreover, the model has a moving con-
tact point on the surface of each link for contact with the ground
surface instead of just a fixed point for each link. The latter is an
approach employed in prior publications on mathematical mod-
els of 3-D snake robot motion. Stick-slip transitions (based on a
set-valued Coulomb friction law) and impacts with the ground
are modeled as instantaneous transitions. This results in an accu-
rate model of spatial Coulomb friction where both the direction
of the friction force and a true stick-phase are taken properly into
account. For wheel-less snake robots, it is important to describe
the frictional contact between the wheel-less snake robot and
the ground in an accurate manner, both with respect to stick-slip
transitions and the direction of the friction force while sliding
along the ground surface. This latter property also distinguishes
wheel-less snake robots from, e.g., legged mechanisms that most
often try to “stick” to the ground rather than sliding along it.
The dynamics of the snake robot is described by an equality
of measures, which includes the Newton–Euler equations for
the nonimpulsive part of the motion as well as impact equa-
tions. A particular choice of coordinates results in an effective
way of writing the system equations. The set-valuedness of the
force laws allows us to write each constitutive law with a sin-
gle equation and avoids explicit switching between equations
(for example, when a collision between the snake robot and the
ground surface occurs) even though this is a hybrid system. This
is advantageous since the snake robot links repeatedly collides
with the ground surface during, e.g., locomotion by sidewinding.
A discretization of the equality of measures gives the so-called
time-stepping method (see [22] and references therein) that we
use for numerical simulation. The description of the model and
the method for numerical integration are presented in this paper
in such a way that people who are new to the field of nonsmooth
dynamics can use this paper as an introduction to nonsmooth
modeling of robot manipulators with impacts and friction. In
addition, we present experimental results that validate the math-
ematical model. To the best of our knowledge, this is the first
time such a back-to-back comparison between simulation and
experimental results is presented for 3-D snake robot motion.
The experiments are performed with the snake robot “Aiko” in
Fig. 1 built at the Norwegian University of Science and Tech-
nology (NTNU)/SINTEF Advanced Robotics Laboratory.
IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 2, APRIL 2008
The paper is organized as follows. A short introduction to the
modeling procedure is given in Section II. The kinematics of the
snake robot with the ground surface as a unilateral constraint
is described in Section III. Then, the groundwork for finding
the friction and ground contact forces is laid in Section IV. The
nonsmooth dynamics is presented in Section V, while the ser-
pentine motion patterns employed in this paper are described in
Section VI. The numerical treatment of the mathematical model
is given in Section VII. Simulations and experimental valida-
tions are given in Section VIII. Conclusions and suggestions for
further research are presented in Section IX.
II. SUMMARY OF THE MATHEMATICAL MODEL
This section contains a brief outline of how to derive the nons-
mooth mathematical model of the snake robot. This preliminary
section is meant to motivate and ease the understanding of the
forthcoming deduction of the system equations.
The snake robot model consists of n links connected by n − 1
two-DOF cardan joints (i.e., rotational joints). Let u ∈ R6n be
a vector containing the translational and rotational velocities of
all the links of the snake robot (the structure of the snake robot
together with the coordinates and reference frames are described
further in Section III). Let the differential measures du and dt
be loosely described for now as a “possible differential change”
in u and time t, respectively, while a more precise definition
is given in Section V. The use of differential measures allows
for instantaneous changes in velocities that occur for impacts
between the snake robot and the ground surface. The system
equations for the snake robot can now be written as
M du − h dt − dR = τ C dt (1)
which is called the equality of measures [23], where M ∈
R6n ×6n is the mass matrix, h ∈ R6n consists of the smooth
forces, τ C ∈ R6n contains all the joint actuator torques, and
dR ∈ R6n accounts for the normal contact forces/impulses
from the ground, the Coulomb friction forces/impulses, and
the bilateral constraint forces/impulses in the joints. Note: We
allow in this paper for instantaneous changes in velocities
usually associated with collisions. Hence, the (normal con-
tact/friction/constraint) forces are not always defined due to the
infinite accelerations. In these cases, we have impulses instead
of forces. The nonsmooth equality of measures (1) allows us to
formulate in a uniform manner both the smooth and nonsmooth
phases of motion. This is achieved partly by representing the
contact forces/impulses as contact impulse measures.
A substantial part of the beginning of this paper is devoted
to deducting the force measure dR. Hence, let us briefly look
at how to derive the contribution of the normal contact impulse
measure between the ground and the first link in dR. Let dR1 ∈
R6 be the sum of contact impulse measures that directly affects
link 1 (i.e., the six top elements of dR), then


friction and joint constraint
dR1 = wN dPN + (2)
impulse measures
where dPN ∈ R is the normal contact impulse measure from
the ground on link 1, and wN ∈ R6 is the corresponding
TRANSETH et al.: 3-D SNAKE ROBOT MOTION: NONSMOOTH MODELING, SIMULATIONS, AND EXPERIMENTS 363

generalized force direction, i.e., a Jacobian (subscripts “j” and

“1” are omitted for brevity).
Let gN ∈ R be a function giving the shortest distance between
the rear part of link 1 and the ground. Such a function is called
a gap function [24]. The gap function is the starting point for
the systematic approach of finding the impulse measures. The
ground and the rear part of the link are separated if gN > 0, are
in contact if gN = 0, and are penetrating each other if gN < 0.
Now, the relative velocity between link 1 and the ground along
the shortest line between the two objects can be defined as
γN := wT
N u1 (3) Fig. 2. Reference frames.

where u1 is the velocity of link 1 and wN = ∂gN /∂q is the

generalized force direction used in (2). It holds that γN = ġN The position and orientation of link i are described by the
for almost all t. The normal contact impulse measure dPN is nonminimal absolute coordinates [25]
related to the relative velocity γN through a set-valued force  
I rG i
law (see Section V-B1. The set-valuedness of the force laws qi = ∈ R7 (4)
allows us to write each constitutive law (force law) with a single pi
equation and avoids explicit switching between equations (for where I r G i = [xi yi zi ]T ∈ R3 is the position of the center
example, when a collision occurs). In addition, this formulation of gravity of link i relative to frame i and the vector pi =
provides an accurate description of the planar Coulomb friction [ei 0 eT T T
i ] , where ei = [ei 1 ei 2 ei 3 ], contains the four Euler
(see Section V-B2). In the following three sections, we will parameters used to describe the rotation. The Euler parameters
elaborate on how to derive the elements that constitute (1), that form a unit quaternion vector with the constraint pT i pi = 1.
is, the various gap functions, relative velocities, and finally, the The coordinates are nonminimal because each link is described
forces/impulses that appear in the equality of measures. with six coordinates, and absolute because the position and
orientation of link i is given directly relative to frame I. The
III. KINEMATICS velocity of link i is given by
The kinematics describes the geometrical aspect of motion.  
I vG i
From the geometry of the snake robot, we develop gap functions ui = ∈ R6 (5)
for ground contact detection. These functions are also needed ω
Bi I Bi

for calculating the directions of the contact forces. where I v G i is the translational velocity of the center of gravity
This section will first give an overview of the coordinates of link i that is I v G i = I ṙ G i when it exists (i.e., for impact-free
used to describe the position and orientation of the snake robot. motion). Moreover, B i ω I B i is the angular velocity of frame
Subsequently, the gap functions will be presented. Bi relative to frame I, given in frame Bi . The transformation
I I
I r = RB i B i r can be performed with the rotation matrix RB i =
A. Coordinates and Reference Frames T
Hi H̄i , where
The snake robot model consists of n cylindrical links that Hi = [−ei ẽi + ei 0 I ] , H̄i = [−ei −ẽi + ei 0 I ] (6)
are connected by n − 1 cardan joints, each having 2 DOFs.
The distance Li between two adjacent cardan joints equals the and the superscript ∼ denotes the skew-symmetric form of a
length of link i, and the radius of each link is LS C i . Each link vector throughout this paper, i.e.,
is modeled as a cylinder of length 2LG S i with two spheres of  
0 −ei 3 ei 2
radius LS C i attached to the ends the link. Link i with parts of  
its two adjacent links are illustrated in Fig. 2. ẽ =  ei 3 0 −ei 1  . (7)
We denote an earth-fixed coordinate frame I = (O, eIx , −ei 2 ei 1 0
ey , eIz ) (see Fig. 2) as an approximation to an inertial frame
I
The time derivative of the rotation matrix is found from [26] as
where its center O is fixed to the ground surface and the eIz -
axis is pointing in the opposite direction of the acceleration ṘIB i = RIB i B i ω̃ I B i = I ω̃ I B i RIB i . (8)
of gravity vector g = [ 0 0 −g ]T . We denote a body-fixed
The coordinates (positions and orientation) and velocities of
frame Bi = (Gi , eIx , eIy , eIz ), where Gi is the center of grav-
all links are gathered in the vectors q = [ q T
1 · · · qT
n]
T
and
ity of link i (which coincides with the geometric center of the
u = [ u1 · · · u n ] .
T T T
link-cylinder) and eIz points along the center line of link i to-
ward link i + 1. A general notation used throughout this paper
is that a vector from the origin of frame I to a point A is B. Gap Functions for Unilateral Constraints
given by r A and a vector from point A to point B is writ- Gap functions for the unilateral constraints (i.e., the ground
ten as r A B . Let a vector r described in frame I be written surface) give the minimal distance between the floor and the
as I r. front and rear part of each link. The contact surfaces between a
364
Fig. 3. Surfaces (solid-drawn circles) on snake robot that constitute the contact
between the robot and the ground.
link and the ground are modeled as two spheres at the ends of
the link, as illustrated for a three-link robot in Fig. 3.
We denote the distance between the center of the two spheres
that belong to link i by 2LG S i and the radius of the spheres by
LS C i . The position of the center of the front and rear spheres are
denoted by r S F i and r S R i , respectively. The shortest distance
between the ground and the points on the front and rear spheres
closest to the ground are denoted by gN F i and gN R i , respectively.
The distances are found from
gN F i = (r S F i )T eIz − LS C i , gN R i = (r S R i )T eIz − LS C i (9)
z , r S R i = r G i − LG S i ez .
where r S F i = r G i + LG S i eB i Bi
The gap functions are gathered in the vector
g N = [ gN F 1 ··· gN F n gN R 1 ···
C. Gap Functions for Bilateral Constraints
Each 2 DOF cardan joint introduces four bilateral constraints,
which will be described by gap functions.
To find the translational “gap” in the joints, we need to
relate the position of the joint between link i and i + 1 to
both links. Let the position of the joint between link i and
i + 1 be written as I r J F i = I r G i + (1/2)Li I eBz , I rJ R i + 1 =
i
Bi+ 1
I r G i + 1 − (1/2)Li I ez . The gap functions can now be found
from
IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 2, APRIL 2008
IV. CONTACT CONSTRAINTS ON VELOCITY LEVEL
In this section, we calculate relative velocities between the
snake robot and the ground from the gap functions. The relative
velocities are needed to set up the set-valued contact forces for
the closed contacts [27].
A. Unilateral Contact: Ground Contact
Contact between the snake robot and the ground involves
(vertical) normal forces, which guarantee the unilaterality of
the contact, and (horizontal) tangential contact forces, which
are due to friction and are dependent on the normal contact
forces and the relative sliding velocities.
1) Relative Velocities Along eIz : The relative velocities be-
tween the front and rear part of link i and the ground along the
eIz -axis are defined as γN F i := ġN F i and γN R i := ġN R i (when
they exist), respectively, and they are used later to find the nor-
mal contact forces. Before we proceed, note that γN F i (or γN R i )
should not be found directly by taking the time derivative of
the expression for gN F i (or gN F i ) in (9). This is the case since
the expressions (9) have already been simplified due to the fact
that expressions have been inserted for the various body-fixed
vectors that constitute the gap functions. Instead, for the relative
velocities, we must consider the velocity of a body-fixed point
PF i that, at time instant t, coincides with a point CF i on the
front sphere closest to the ground (the same principle applies
gN R n ]T . (10) for the rear part of the link). Note that the position vectors for
PF i and CF i will be the same instantaneously. However, the
differentials will be different. This is shown in the following.
Let CF i be a point on the front sphere that moves on the sphere
such that the point is always closest to the ground surface. Then
the position of this point is
rC F i = rG i + rG i S F i + rS F i C F i . (14)
Define the skew-symmetric matrix I r̃ G i C F i = I r̃ G i S F i +
The velocity of the point CF i is obtained by dif-
I r̃ S F i C F i .
ferentiation of (14) as
I v C F i =I v G i −I r̃ G i C F i RB i B i ω I B i +RB i B i ṙ S F i C F i
I I
(15)
 T
 
gJ i χ = I rJ F i − I rJ R i + 1 I
I eχ (11) I vP F i

where we now can use

for i = 1, . . . , n − 1, and χ = x, y, z. Hence,
 T r S F i C F i = −LS C i eIz (16)
gJ i χ = I rG i − I rG i + 1 I
I eχ
and we have employed (8) and B i ṙ G i S F i = 0, together
1 
Bi+ 1 T with the identities I ω̃ I B i I r G i S F i = − I ω I B i I r̃ G i S F i and
+ Li RIB i Bi I
B i ez +Li+1 RB i + 1 B i + 1 ez
I
I eχ (12)
2 I ω̃ I B i I r S F i C F i = − I ω I B i I r̃ S F i C F i . We see from (15) an ex-
pression for the velocity I v P F i of the body-fixed point PF i that,
for i = 1, . . . , n − 1. at time instant t, coincides with the point CF i . This velocity will
The next gap function provides a “gap” in rotation around the be employed to find both the relative velocities along eIz and the
axis that a cardan joint is not able to rotate around. Let eB
y and
i
tangential relative velocities. The equivalent velocity I v P R i on
Bi+ 1
ex be the axes of rotation for the cardan joint between link the rear part of the sphere is found similarly to I v P F i in (15) by
Bi+ 1
i and link i + 1, then (eB i T
y ) ex = 0. Hence, a measure for replacing F i by Ri in (15) and (16).
the rotational “gap” can be defined from the previous equality Now, the relative velocities along eIz for the front and rear
as the gap function part of link i can be found as
   I
Bi T
 γN Q i = ( I eIz )T I v P Q i =⇒ γN Q i = (wN Q i )T ui (17)
gJ i φ = RIB i B i ey RB i + 1 B i + 1 eB
x
i+ 1
. (13)
TRANSETH et al.: 3-D SNAKE ROBOT MOTION: NONSMOOTH MODELING, SIMULATIONS, AND EXPERIMENTS 365

where Until now, the relative velocities have been given in the di-
rections along the eIx - and eIy -axes. In order to calculate the
wN Q i = [ (I eIz )T −(I eIz )T I r̃ G i C Q i RIB i ]T (18) friction forces longitudinal and transversal to a link, we need
for Q = F, R with r G i S F i = LG S i eB z
i
and r G i S R i = to know the corresponding relative velocities in these directions
−LG S i ez . The motivation to use the form (17) is that wN Q i
Bi in the (eIx , eIy )-plane. To calculate these velocities, we intro-
gives the generalized direction of the contact force between the duce for each link a frame Πi with axes (eΠ Πi Πi
x , ey , ez ), where
i

ground and the front and rear part of link i. Πi I

I ez = I ez , and
A vector gathering all γN F i and γN R i is
Πi Axy I eB
z
i
Πi
I
I ez × I eΠ
x
i
T
γN = WN u (19) I ex = , I ey = (27)
Axy I ez 
Bi
 I eIz × I ex 
Πi

where [γN = [ γN F 1 · · · γN F n γN R 1 · · · γN R n ]T , WN = where Axy = diag([1, 1, 0]). Hence, it holds that

[ WN F WN R ] ∈ R6n ×2n , and
  RIΠ i = [ I eΠ
x
i Πi
I ey
Πi
I ez ]. (28)
wN Q 1 06×1 · · · 06×1
 ..  Notice that I eΠ x = I ez
i Bi
and I eΠ i Bi
y = I ex when link i is
 0 . .. 
 6×1 .  Bi I
lying flat on the ground with I ey = I ez . Since only the motion
WN Q =   (20)
 . . ..  in the (eIx , eIy ) -plane is of interest, we define
 . . 06×1 
 T
06×1 ··· 06×1 wN Q n 1 0 0
R̄IΠ i = DT RIΠ i D, D= . (29)
for Q = F, R. 0 1 0
2) Tangential Relative Velocities: Friction forces between a
Since we now have that γ T Q i = R̄IΠ i γ T Q i , Q = F, R, the rela-
link and the ground depend on the relative velocities in the
tive velocity between the floor and the front part of link i, with
(eIx , eIy )-plane between the snake robot links and the ground.
respect to frame Πi , can be found from (25) as
These velocities are termed tangential relative velocities. First,
the relative velocities between the front part of link i and the γT Q i = WTTQ i ui (30)
ground along the eIx - and eIy -axis, γT F i x and γT F i y , will be
deducted. Subsequently, the relative velocities for the front part where γ T Q i = [ γ T Q i x γ T Q i y ]T , WT Q i = WT Q i R̄IΠ i , for
of link i along the projection of the longitudinal axis of the link Q = F, R.
onto the (eIx , eIy )-plane γT F i x and transversal to the link γT F i y , A vector that gathers all γ T F i and γ T R i is found from
will be deducted from γT F i x and γT F i y , respectively. The same
γ T = WTT u (31)
type of notation applies for the rear part of link i: γT R i x , γT R i y ,
γT R i x , and γT R i y . The tangential relative velocities are found where
much in the same way as for γN F i and γN R i . Consequently, we T
find the velocities of the points PF i and PR i along the eIx - and γT = [ γT
TF 1 ··· γT
TF n γT
TR 1 ··· γT
TR n ] (32)
eIy -axis. Hence, by looking at (17), we see that the tangential WT = [ WT F WT R ] ∈ R6n ×4n , and WT F , WT R are found
relative velocities of the contact points on the front part of the similarly to (20) by replacing the zero vectors with 06×2 and
link can be found as replacing the vectors wN F i , wN R i with the matrices WT F i ,
γ
TF i ζ= ( eI )T v
I ζ I =⇒ γ 
PF i = (w
TF i ζ )T u TF i ζ (21) i
WT R i , respectively.
3) Relative Rolling Velocities: Up until now, we have only
for ζ = x, y, where considered the translational relative motion of the snake robot
links. However, we also need to consider rotational relative
wT F i ζ = [ (I eIζ )T −(I eIζ )T I r̃ G i C F i RIB i ]T . (22)
motion to add a damping effect on the rotational motion in the
The relative velocities between the rear part of the link and form of rotational friction. In order to do this, we introduce a
the ground are found by exchanging I v P F i with I v P R i in (21) relative rolling velocity as
 
γT R i ζ = (I eIζ )T I v P R i =⇒ γT R i ζ = (wT R i ζ )T ui (23) γ V Q i = DT I r C Q i S Q i × I ω I B i (33)
for ζ = x, y, where for Q = F, R, where γ V Q i = [ γV Q i x γV Q i y ]T and
I
wT R i ζ = [ (I eIζ )T −(I eIζ )T I r̃ G i C F i RIB i ]T . (24) I r C Q i S Q i = LS C i I ez is the vector pointing (upwards) from
the body-fixed point CQ i on the link end sphere momentarily
The tangential relative velocities for the front and rear part of closest to the ground, toward the center SQ i of the end sphere.
the link i are combined in vectors By employing the identity
γT Q i = WTTQ i ui (25) I rC Q i S Q i × I ω I B i = I r̃ C Q i S Q i RIB i B i ωI B i (34)
for Q = F, R, where γT Q i = [ γT Q i x γT Q i y ]T and we find that the relative rolling velocity can be written as
 
WT Q i = wT Q i x wT Q i y ∈ R6×2 . (26) γV Q i = WVTQ i ui (35)
366 IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 2, APRIL 2008

J 1 · · · γ J n −1 ] ,
where γ J = [ γ T
where T T

 T
WVTQ i = [ O2×3 DT I r̃ C Q i S Q i RIB i ] (36) WJT1 04×6 ··· 04×6
 .. 
0 WJT2 
for Q = F, R.  4×6 . 
WJ =   ∈ R6n ×4(n −1)
We gather all the relative rolling velocities as  .. .. 
 . . 04×6 
γ V = WV u (37) 04×6 ··· 04×6 WJTn −1
(43)
where γ V = ···
[ γT γT ··· γT γT T and WJ i = [ wJ i x wJ i y wJ i z wJ i φ ] ∈ R12×4 , for i =
VF 1 VF n VR 1 WV = VR n ] ,
[ WV F WV R ] ∈ R6n ×4n , and WV F , WV R are found simi- 1, . . . , n − 1.
larly to (20) by replacing the zero vectors with 06×2 and replac-
ing the vectors wN F i , wN R i with the matrices WV F i , WV R i , V. NONSMOOTH DYNAMICS
respectively. The starting point for describing the dynamics of the snake
robot is the equality of measures as introduced in [23]. The
B. Bilateral Constraints: Joints equality of measures includes equations of motion for impact-
free motion as well as impact equations. The impact equations
Bilateral constraints introduced by the cardan joint between
give rise to impulsive behavior [24]. In this section, we employ
two adjacent links prohibit relative motion between the links in
the results from Sections III and Section IV in order to find the
4 DOFs. The relative velocities between the links along these
equality of measures for the snake robot.
DOFs need to be found in order to calculate the bilateral con-
straint forces in the joints. These relative velocities are found
from the gap functions (12) and (13). A. Equality of Measures
Relative velocities for the translational gap between link i An equality of measures describes the dynamics of the snake
and link i + 1 are defined as γJ i χ := ġJ i χ , for i = 1, . . . , n − 1, robot within the context of nonsmooth dynamics. Velocity
where χ = x, y, z. By employing (12), we find that jumps, usually associated with impacts, are modeled to occur
  instantaneously. Hence, the time derivative of a velocity does
ui not always exist. By considering the generalized velocity to be
γJ i χ = wT
Ji χ (38)
ui+1 a function t → u(t) of locally bounded variation on a time in-
terval I = [ tA , tE ] [23], the function u(t) admits a right u+
where and left u− limit for all t ∈ I, and its time derivative u̇ exists
  for almost all t ∈ I. To be able to obtain u from integration, we
(I eIχ )
    need to use the differential measure du, where it is assumed that
 − (I eIχ )T L2i RIB i Bi T
B i ẽz
 the measure can be decomposed into
 
wJ i χ =

 . (39)
  
 −(I eIχ )  du = u̇ dt + u+ − u− dη (44)
  T 
− (I eIχ )T (Li+1 /2)RIB i + 1 Bi+ 1
B i + 1 ẽz where dt denotes the Lebesgue
 measure and dη denotes the
atomic measure where {t 1 } dη = 1. The total increment of u
A relative velocity for the rotational gap is defined as γJ i φ := over a compact subinterval [t1 , t2 ] of I is found as
ġJ i φ , for i = 1, . . . , n − 1. Hence, 
  du = u+ (t2 ) − u− (t1 ) (45)
ui [t 1 ,t 2 ]
γJ i φ = wT
Ji φ (40)
ui+1 and is due to a continuous change (i.e., impact-free motion)
stemming from u̇ as well as possible discontinuities in u due
where to impacts within the time interval [t1 , t2 ]. Equation (45) is also
  valid when the time interval reduces to a singleton {t1 }, and
03×1
  I   I
Bi T Bi+ 1   if a velocity jump occurs for t = t1 , then (45) gives a nonzero
 − RB i B i ẽy RB i + 1 B i + 1 ex 
wJ i φ =

.
 (41) result.
 03×1  From the earlier notation, the Newton–Euler equations as an
   I
Bi+ 1 T
 equality of measures can be written in a general form as
− RIB i + 1 B i + 1 ẽx RB i B i eB
y
i

M (q, t) du − h (q, u, t) dt − dR = τ C dt (46)

Let γJ i = [ γJ i x γJ i y γJ i z γJ i φ ]T , then we can gather all
the relative velocities concerned with the bilateral constraints in where the mass matrix M(q, t), the vector of smooth forces
one vector h(q, u, t), the force measure of possibly atomic impact impul-
sions dR, and the vector of applied torques τ C will be described
γ J = WJT u (42) in the following.
TRANSETH et al.: 3-D SNAKE ROBOT MOTION: NONSMOOTH MODELING, SIMULATIONS, AND EXPERIMENTS 367

For our choice of coordinates, the mass matrix is diagonal where λN a is the Lebesgue-measurable force and ΛN a is the
and constant purely atomic impact impulse. The same decomposition can
  also be performed for the other three impulse measures.
M1 0
The control torques τ C are described in Section V-C.
 ..  6n ×6n
M(q, t) = M =  . ∈R (47)
0 Mn B. Constitutive Laws for the Contact Forces
In this section, we will introduce set-valued force laws for
where
normal contact and Coulomb friction. These laws will all be
 
mi I3×3 03×3 formulated on velocity level using the relative contact velocities
Mi = (48) γ given by (19) and (31). Subsequently, the set-valued force
03×3 B i ΘG i
laws are formulated as equalities in Section V-B.4 using the
with B i ΘG i = diag([Θ1i Θ1i Θ3i ]), mi is the mass of so-called proximal point function in order to include the force
link i, and Θ1i and Θ3i are its moments of iner- laws in the numerical simulation [22].
tia. The smooth forces, here consisting of gravity and 1) Normal Contact Force: The normal contact between a
gyroscopic accelerations, are described by h(q, u, t) = link and the floor is described by the unilateral constraint
h(u) = [hT 1 (u1 ) · · · hTn (un )] ∈ R , where hi (ui ) =
T 6n

[0 0 − mi g − (B i ω̃ I B i B i ΘG i B i ω I B i )T ]T .
The force measure dR accounts for all the contact forces and
impulses. The contact efforts that constitute dR are found from
the force laws given in Section V-B. Let I be the set of all active
contacts with the ground
I(t) = {a | gN a (q(t)) = 0} ⊆ {1, 2, . . . , 2n}
where gN a is the ath element of the vector g N in (10). Now, the
force measure is written as
 express all force laws for closed contact on velocity level, while
dR = WJ dP J + (WN )a dPN a
a∈I
gN ≥ 0, λN ≥ 0, gN λN = 0 (52)
which is known as Signorini’s law [22]. Here, λN is the normal
contact force and gN is the gap function. Subscripts Ri and F i
are temporarily removed for simplicity. This set-valued force
law states that the contact is impenetrable, gN ≥ 0, the con-
tact can only transmit pressure forces λN ≥ 0, and the contact
(49)
force λN does not produce work gN λN = 0. The force law can
be expressed on different kinematic levels: displacement level
(52), velocity level, and acceleration level. In the following, we
all forces vanish for open contacts. Then, by employing con-
cepts of convex analysis, the relationship between the relative
  velocity and the Lebesgue-measurable normal contact force (not
+ (WT )2a−1 dPT a x + (WT )2a dPT a y
an impulse) may be written for a closed contact gN = 0 as an
a∈I
  inclusion on velocity level
+ (WV )2a−1 dPV a x + (WV )2a dPV a y
a∈I
where dPN a is the normal contact impulse measure between
the ground and a link, dPT a x and dPT a y are the tangential con-
tact impulse measures (friction) between the floor and a link,
longitudinal and transversal to the link (i.e., along the eΠ
Πi
ey -axis), respectively, dPV a x and dPV a y are the rotational con-
tact impulse measures (friction) between the floor and a link,
along the eIx -axis and eIy -axis, respectively, dP J is the contact
impulse measure due to the bilateral constraints in the joints
(these constraints are always active), and the lower case sub-
scripts on the W matrices indicate which column of the matrix
is used. The position of the elements of the vectors dP N , dP J ,
dP T , and dP V corresponds to the position of the elements in
their respective vector of relative velocity γ N , γ J , γ T , and γ V .
Hence, looking at, for example, the expression (32) for γ T , we
see that, e.g., dP T n + 1 = [dPT ( n + 1 ) x dPT ( n + 1 ) y ]T corresponds
to γ T R 1 , and we know from this that dP T n + 1 is the tangential
contact impulse measure between the ground and the rear part
of link 1.
The contact impulse measures can be decomposed in the same
way as for du. Let us take the normal contact impulse measure
as an example. The measure can be written as
dPN a = λN a dt + ΛN a dη
(50)
−γN ∈ NC N (λN ) (53)
where the convex set CN = {λN | λN ≥ 0} = R is the set+
of admissible contact forces, and NC N is the normal cone to
CN [22]. The inclusion (53) is equivalent to the condition
x - and
i
γN ≥ 0, λN ≥ 0, γN λN = 0 (54)
for a closed contact gN = 0. Before explaining the force law
(53), let us first mention that this force law describes the impen-
etrability of sustained contact, i.e., gN = 0 and γN = 0, as well
as the detachment: γN > 0 ⇒ λN = 0. However, (53) does not
cover impacts (where we have impulses instead of forces). For
impacts, we need a similar impact law described at the end of
this section.
From the definition [22], [28] of a normal cone NC (x) to a
convex set C at the point x ∈ Rn , we have that NC (x) = {0},
for x ∈ int C, and NC (x) = {∅}, for x ∈ / C. If x is on the
boundary of C, then NC (x) is the set of all vectors y ∈ Rn
that are normal to x. For example, for CN = R+ , we have
NC N (0) = R− and NC N (2) = 0.
The force law (53) only covers finite-valued contact efforts
during impulse-free motion, i.e., all velocities are absolutely
continuous in time. When a collision occurs in a rigid-body
setting, then the velocities will be locally discontinuous in order
(51) to prevent penetration. The velocity jump is accompanied by an
368
Fig. 4. Relationship between tangential relative velocity and friction force.
The set C T is in gray.
impact impulse ΛN , for which we will set up an impact law. The
+
relative velocity admits, similarly to the velocities u, a right γN
− law always slide. Second, the very steep slope of the friction
and a left γN limit. The impact law for a completely inelastic
impact at a closed contact can now be written as
−γN
+
∈ NC N (ΛN ), CN = {ΛN | ΛN ≥ 0} = R+ (55)
which is equivalent to the condition
+
γN ≥ 0, ΛN ≥ 0, +
γN ΛN = 0.
Notice that the force law (53) and impact law (55) is on the same
form. We have earlier stated that there is no need for an explicit
switch between equations when, e.g., an impact occurs. This be-
comes evident in Section VII with the introduction of the contact
impulse (that includes both forces and possible impulses) for the
discretization of the system dynamics.
The force law for normal contact given in this section can also
be employed to describe normal contact with obstacles. This is
described in [29].
2) Coulomb Friction Force: In this section, we describe the
friction force between the snake robot and the ground when
the snake robot slides along the ground surface as set-valued
Coulomb friction. Similarly to the force law (53) for normal
contact, we describe the constitutive description for friction us-
ing an inclusion on a normal cone. The friction force λT , in
the 2-D tangent plane to the contact point, is modeled with an
anisotropic Coulomb friction law
−γ T ∈ NC T (λ
λT )
where γ T is a relative sliding velocity, NC T is the normal cone
to the set CT , and the ellipse
 
λ2T x λ2T y
IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 2, APRIL 2008
conclude that γ T = 0. Obviously, this corresponds to the stick
phase of the friction law. If the friction force is on the boundary
of CT , then the normal cone NC T (λ λT ) consists of the outward
normal ray on the ellipse CT at the point λT .
The advantage to formulate the friction law as the inclusion
(57) now becomes apparent. First of all, note that (57) is a
spatial friction law. Such a spatial friction law cannot properly
be described by a set-valued sign function. Some authors [10],
[12] model the spatial contact with two sign functions for the
two components of the relative sliding velocity using two fric-
tion coefficients µT x and µT y . Others smoothen the (set-valued)
sign function with a smoothening function, e.g., some arctan-
gent function. This results in a very steep slope of the friction
curve near zero relative velocity. Such an approach is very cum-
bersome for two reasons. First of all, stiction cannot properly be
described: an object on a slope will with a smoothened friction
curve causes the differential equations of motion to become nu-
merically stiff. Summarizing, we see that (57) describes spatial
Coulomb friction taking stiction properly into account.
While the force law (57) only describes the Coulomb friction
during impulse-free motion, we also need a force law for impact
(56) impulses λT . These are found from the exact same form as (57)
by replacing γ T with its right limit γ +T and inserting λ T instead
of λT both in (57) and in the convex set CT in (58).
3) Rolling Friction: The snake robot modeled in this paper
has cylindrical links which means that it may roll sideways. The
spatial Coulomb friction described in the previous section arises
from translational motion. In addition, there is also a force that
resists the rolling motion of the snake robot. We model this force
as a “rolling friction” λV ∈ R2 (additional subscripts omitted
for simplicity) in this paper by employing the same set-valued
force law as for the tangential Coulomb friction force λT in
Section V-B2. However, we consider the rolling friction to be
isotropic, and therefore, we find the rolling friction as
−γ V ∈ NC V (λ
λV ) (59)
where
CV = {λ
λV | λ
λV  ≤ µV λN } (60)
and µV > 0 is the friction coefficient.
(57) The general form of the force law (59) is also valid for impact
impulses ΛV in the same way as for the tangential Coulomb
friction described in Section V-B2. Subsequently, the impact
impulse is found by exchanging γ V with γ + V and λ V with ΛV
in (59) and (60).
CT = λT | + 2 ≤ 1 (58)
(µT x λN )2 µT y λN
is the set of admissible friction forces, where λT =
[λT x λT y ]T , and µT x , µT y > 0 are directional friction coeffi-
cients along the eΠ Π
x - and ey -axis from (27), respectively. Fig. 4
depicts the set CT (in gray), together with the relationship be-
tween the tangential relative velocities and the friction force
when it is on the boundary of CT .
The force law (57) distinguishes between two cases: if the
friction force is in the interior of CT or on its boundary. If
λT ∈ int CT , then it holds that NC T (λ λT ) = 0 from which we
4) Constitutive Laws as Projections: An inclusion cannot be
directly employed in numerical calculations. Hence, we trans-
form the force laws (53), (57), and (59) that have been stated as
an inclusion to a normal cone, into an equality. This is achieved
through the so-called proximal point function proxC (x), which
equals x if x ∈ C and equals the closest point in C to x if x ∈
/ C.
The set C must be convex. Using the proximal point function,
we transform the force laws into implicit equalities [22]
−γκ ∈ NC κ (λκ ) ⇐⇒ λκ = proxC κ (λκ − rκ γκ ) (61)
where rκ > 0 for κ = N, T, V .
TRANSETH et al.: 3-D SNAKE ROBOT MOTION: NONSMOOTH MODELING, SIMULATIONS, AND EXPERIMENTS 369

side to side, and αv i describes the lifting and lowering of the

links. The rotation angles can now be found from the relative
rotation matrix (63) as
 Bi 
−1
(RB i + 1 )32
αv i = tan (65)
(RBB i + 1 )33
i

 i 
αh i = −sin−1 RB B i + 1 31 (66)

for i = 1, . . . , n − 1, where (RB

B i + 1 )32 is the element of the
i

matrix RB B i + 1 in row 3 column 2, etc.

i

The rotational velocities of the joints are found directly from

the rotational velocities of the links. We define the rotational
velocity for sideways motion as ωh i = dT2 ω J i and lifting motion
as ωv i = dT1 ω J i for the joint between link i and i + 1, where
dT1 = [ 1 0 0 ], dT2 = [ 0 1 0 ], and
Fig. 5. Control torques. (a) Side view. (b) Top view.
C. Joint Actuators
Each cardan joint has 2 DOF that are controlled by two joint
actuators. The actuators are modeled as controlled torques ap-
plied around the axes of rotation for the joint. Fig. 5 illustrates
how the direction of positive rotation is defined. Define for link i
a positive control torque τv i to give a positive rotational velocity
B
around ex i + 1 and a positive control torque τh i to give a positive
rotational velocity around eB y , both with respect to link i. The
i
 i T
ω J i = B i + 1 ω I B i + 1 − RB
B i + 1 B i ωI B i . (67)
Let the desired values of αh i and αv i be αh i ,r and αv i ,r ,
respectively. In addition, the reference values for the joint ve-
locities ωh i , ωv i are given by α̇h i ,r and α̇v i ,r , respectively. Then,
proportional derivative (PD) controllers for the joints are
τh i = Kh p (αh i − αh i ,r ) + Kh d (ωh i − α̇h i ,r ) (68)
τv i = Kv p (αv i − αv i ,r ) + Kv d (ωv i − α̇v i ,r ) (69)

total torque τ C i ∈ R applied to link i is

3
for i = 1, . . . , n − 1, where Kh p , Kh d , Kv p , and Kv d are posi-
        tive constants and equal for all i.
0 0 τv i τv ( i −1 )
       
B i −1  τh ( i −1 )  + RB i + 1  0  − 
τ C i =  τh i  − RB i Bi
0 
B. Motion Pattern and Reference Angles
0 0 0 0
(62) A general expression for defining some of the most common
for i = 1, . . . , n, where the relative rotation matrix is motion patterns for snake robots is given by
 I T I
RB
B i + 1 = RB i
i
RB i + 1 (63) αh i ,r = Ah sin (ωh t + (i − 1)δh ) + ψh (70)
and τh 0 = τv 0 = τh n = τv n = 0. The vector of the torques ap- αv i ,r = Av sin (ωv t + (i − 1)δv + δ0 ) + ψv (71)
plied to all links τ C ∈ R6n is
T
for i = 1, . . . , n − 1, where αh i ,r and αv i ,r are the reference
τ C = [ 01×3 τT
C1 01×3 C 2 · · · 01×3
τT τT
C n ] . (64) angles for αh i and αv i , and Ah , Av are the amplitude of oscil-
lation, ωh , ωv are the angular frequencies, δh , δv are the phase
VI. MOTION PATTERN AND JOINT CONTROL offsets, and ψh , ψv are the angle offsets, for the horizontal and
In this section, we will define the joint angles and show how vertical waves, respectively [3]. The offset between the vertical
to control them for snake robot locomotion. and horizontal waves is given by δ0 .
Two motion patterns that biological snakes use have been em-
A. Accessing and Control of Joint Angles ployed in this paper. The first motion pattern is called “lateral
undulation” (also denoted “serpentine crawling”) where loco-
The joint angles are not directly accessible from the nonmini- motion is obtained by propagating horizontal waves from the
mal coordinates, but can be calculated from the relative rotation front to the rear of the snake body while exploiting roughness
matrices RB Bi
B i + 1 in (63). Assume that RB i + 1 is constructed from
i
in the terrain and digging its body into the ground. These latter
successive rotations (Euler angles with the zyx-convention) αz i , two properties are the motivation for our anisotropic friction
αh i , and αv i : RB B i + 1 = Rα z i Rα h i Rα v i . Since we have cardan
i
model. The second motion pattern is called “sidewinding” and
joints, let Rα z i = I3×3 be the rotation around the eB z -axis, and
i
is mainly used by biological snakes moving on uniform sur-
Bi B
let Rα h i and Rα v i describe the rotation around ey and ex i + 1 , faces [30]. Parts of the snake are lying relatively stationary on
respectively. Hence, αh i describes the DOF of the cardan joint the ground while the rest of the body is lifted and moved forward
between link i and link i + 1 that moves link i and i + 1 from resulting in a looping movement during locomotion.
370 IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 2, APRIL 2008

VII. NUMERICAL ALGORITHM—TIME STEPPING Section V-B4 for the set of active contact points I. Hence, the
The numerical solution of the equality of measures is found constitutive laws (61) for the ground contact impulsions may
with an algorithm introduced in [23] (see also [22] and [24]) now be written as
called the time-stepping method described in the following. The PN a = proxC N (PN a − rN γN E a ) (77)
methods applied in Section VII-A and Section VII-B are based  
on [27], except for the direct calculation of the bilateral contact P T a = proxC T P T a − rT γ T E a (78)
impulsions that we have introduced.  
P V a = proxC V P V a − rV γ V E a (79)
A. Time Discretization where rN , rT , rv > 0, a ∈ I, γN E a is the ath element of the
Let tl denote the time at time step l = 1, 2, 3, . . . , where the vector γ N E , and γ T E a , γ V E a are the vectors of the (2a − 1)th
step size is ∆t = tl+1 − tl . Consider the (usually very short) and 2ath elements of γ T E and γ V E , respectively, and
time interval I = [tA , tE ] and let tA = tl . Define q A = q(tA ), T
γ N E = γ N (q M , uE ) = WN M uE (80)
uA = u(tA ) that are admissible with respect to both the unilat-
eral and bilateral constraints, and the unit norm constraint on the γ T E := γ T (q M , uE ) = WTT M uE (81)
Euler parameters. Our goal is now to find q E = q(tE ). We use γ V E := γ V (q M , uE ) = WVT M uE . (82)
the states of the system at the mid point tM = tA + (1/2)∆t
of the time interval I to decide which contact points are ac- The constitutive laws (77)–(82) are valid for completely in-
tive (i.e., which links are touching the ground). The coordinates elastic impact.
(positions and orientations) at tM are found from The constraints on the joints are bilateral, and it therefore,
∆t holds that γ J E := γ J (q E , uE ) = WJTM uE = 0 ∀ t. This al-
qM = qA + F(q M )uA (72) lows us to directly compute the associated contact impulsions
2
P J by solving (75) for P J with uE = 0. By solving for P J
where we have used the equality ṗ = (1/2)H̄T B i ω I B i [26], and solving (75)–(81), we find that
 
FH 1 07×6 · · · 07×6 ∆t
 ..  qE = qM + F(q M )uE . (83)
0 ··· .  2
 7×6 FH 2 
F(q M ) =   (73)
 .. ..  B. Solving for the Contact Impulsions
 . . 07×6 
07×6 · · · 07×6 FH n In this section, we show how to calculate the contact
  impulsions PN a , P T a , P V a , and P J in (76) for a ∈ I. The
I3×3 03×3 numerical integration algorithm used in this paper is called a
1 T ∈R
7×6
FH i = (74)
04×3 H̄i time-stepping method that allows for a simultaneous treatment
2 of both impulsive and nonimpulsive forces during a time step.
and H̄i is found from (6) by inserting the orientation of link The frictional contact problem, defined by (75)–(82) and finding
i at time tM . The approximation of the matrices WΞ , where P J , needs to be solved for each time step tl . A modified Newton
Ξ = N, T, J, V , on the time interval I is given as WΞM := algorithm [31] has been chosen to solve the nonlinear problem
WΞ (q M ). The same applies for hM := h(q M , uA ). A numer- iteratively because of its simplicity. Let the superscript (k )
ical approximation of the equality of measures (46) over the denote the current iteration of the modified Newton algorithm,
time interval I can now be written as and initialize all contact impulsions (for active contacts) with
the value they had the last time their corresponding contact point
M (uE − uA ) − hM ∆t − S − WJ M P J = τ C ∆t (75)
was active. Let those that were active be initialized with their
where previous values. Now, the algorithm may be written as follows.
 1) Solve
S= (WN M )a PN a + [(WT M )2a−1 (WT M )2a ] P T a
 −1
P J = WJTM M−1 WJTM
a∈I (k )
   
+ [(WV M )2a−1 (WV M )2a ] P V a (76)
· WJTM uA −WJTM M−1 hM ∆t+S (k ) +τ C ∆t (84)
a∈I

and PN a , P T a , P V a , and P J are the contact impulsions dur- where

ing the time interval I. They consist of forces λ acting during  (k )
I, and possible impulses Λ acting in the time interval I. The S (k ) = (WN M )a PN a
subscript a denotes which link the contact impulsions are acting a∈I

on, and is employed the same way as for the contact impulse + [(WT M )2a−1 (WT M )2a ] P T a
(k )
measures in (50). To find the positions and orientations q E at a∈I
the end of the time interval, we need to solve (75) for uE and  (k )
the contact impulsions. The contact impulsions associated with + [(WV M )2a−1 (WV M )2a ] P V a . (85)
ground contact are found using the prox functions described in a∈I
TRANSETH et al.: 3-D SNAKE ROBOT MOTION: NONSMOOTH MODELING, SIMULATIONS, AND EXPERIMENTS 371

2) Solve uE
(k +1)
from given with orthotropic friction. Finally, the latter three sections
 (k +1)  present the experimental setup together with simulation and ex-
(k )
M uE −uA −hM ∆t − S (k ) −WJ M P J = τ C ∆t. perimental results of the motion patterns lateral undulation and
(86) sidewinding with isotropic friction.
3) Solve for a ∈ I
(k +1)  (k ) (k +1)  A. Model and Simulation Parameters
P Na = proxC N PN a − rN γN E a (87)
  The snake robot Aiko in Fig. 1 was used as a basis for the
(k +1) (k ) (k +1)
P Ta = proxC T P T a − rT γT E a (88) parameters in the mathematical model. The model parameters
(k +1)  (k ) (k +1)  are for i = 1, . . . , n: Aiko has n = 11 links. The length Li =
P Va = proxC V P V a − rV γV E a (89) 0.122 m and radius LS C i = 0.0525 m of each link is found
where by measuring one of the links used on Aiko. Moreover, Aiko
(k +1) T (k +1)
weighs 7.5 kg, and we therefore, assume that each link weighs
γN E = WN M uE (90) mi = 7.5/11 kg ≈ 0.682 kg since all links are approximately
(k +1) (k +1) equal. The distance from the center of gravity of a link to the
γT E = WTT M uE (91)
center of the spheres used for contact with the ground in the
(k +1) (k +1)
γV E = WVT M uE . (92) model is calculated to be LG S i = 0.0393 m by assuming a
45◦ maximum joint deflection and that the cylindrical parts of
Repeat steps 1)–3) until the model should not come into each other for a maximum
 (k +1) (k )   (k +1) (k )  joint angle. The moments of inertia Θ1i = 9.63 × 10−4 kg·m2 ,
P − P N  + P T − PT 
N
Θ3i = 2.35 × 10−4 kg·m2 are calculated by assuming each link
(k +1) (k )
+ P V − PV  <
(93) to be a cylinder of length Li and radius LS C i with a uniform
mass distribution.
where
> 0 is a user-defined tolerance. Subsequently, q E is cal- The controller parameters are Kh p = 40 Nm, Kh d = 0.2
culated from (83) and the calculation of the time step is finished. Nm·s, Kv p = 800 Nm, and Kv d = 0.2 Nm·s. The accelera-
Usually, a higher value of the parameters rN , rT , rV yields a tion of gravity is g = 9.81 m/s2 . The simulation parameters
faster convergence rate at the risk of divergence. However, a are as follows: rN = 0.1, rT = 0.01, rV = 0.05, tstart = 0
general convergence result for the modified Newton algorithm s, tstop = 15 s, ∆t = (tstop − tstart )/N seconds, and N =
does not exist. The constitutive laws (87)–(92) used to describe 4000(tstop − tstart ). The Coulomb friction coefficients are
the contact impulses are given on velocity level. This means given together with the presentation of the simulations since
that the bilateral constraints on position level, and the unit norm we have not employed the same friction coefficients for all
constraint on the Euler parameters are, in general, not satisfied. simulations.
A solution to these problems is suggested in the following.
B. Lateral Undulation: Simulation Results
C. Constraint Violation
In this section, we let the snake robot move by the motion pat-
After the modified Newton algorithm has converged and q E
tern lateral undulation (see Section VI). We include this motion
has been found from (83), the unit-norm constraint pi  = 1 is
pattern since it is commonly used for snake robots [1], [10]–[13],
satisfied from pnewE i = pE i /pE i , for i = 1, . . . , n, where pE i [32]–[34], and we therefore, want to show that the desired (for-
is the quaternion describing the orientation of link i. The new
ward) motion is obtained for our model. We do not provide an
quaternions should now be inserted into q E .
experimental validation in this section since our snake robot
The links have to be projected so that the bilateral constraints
Aiko does not have the orthotropic friction property necessary
(12), (13) are satisfied. This is done in a two-step process while
for efficient locomotion by lateral undulation on a flat surface.
keeping the position and orientation of link 1 fixed. First, the
However, we compare simulation and experimental results for
orientation of the positions of the remaining links are altered so
lateral undulation with isotropic friction in Section VIII-D.
that gJ i φ = 0 is satisfied for i = 1, . . . , n − 1. In this process, all
The Coulomb friction coefficients µT x = 0.1 (along the snake
the eB Bi Bi
z -axes are kept fixed, while the ex and ey are changed,
i
robot body) and µT y = 0.5 (transversal to the body) are em-
if necessary. Subsequently, the remaining links are translated so
ployed for the simulation of lateral undulation in this section.
that gJ i χ = 0 is satisfied for i = 1, . . . , n − 1, χ = x, y, z. The
Hence, orthotropic friction is obtained. The motion pattern is im-
new positions of the links should now be inserted into q E . Now,
plemented using the joint reference angles given by (70) and (71)
the positions q E and velocities uE can be used as the initial
with the parameters Ah = 30π/180 rad, ωh = 80π/180 rad/s,
states for the next time step tl+1 .
δh = −50π/180 rad, Av = δv = ψv = ψh = δ0 = 0 rad, and
ωv = 0 rad/s. We let the snake robot start in a curved posture
VIII. SIMULATIONS AND EXPERIMENTAL VALIDATION with its center line approximately along the eIx -axis and with its
In this section, we present simulations of the mathemati- initial joint angles given by αh i (0) and αv i (0), respectively.
cal model together with experiments with a real snake robot. Fig. 6 shows the position of the center of gravity of the middle
First, we present the model and simulation parameters. Second, link (i = 6) of the snake robot. We see that the snake robot
simulation results of the motion pattern lateral undulation are behaves as expected: the snake robot starts moving steadily
372
Fig. 6. Simulation results (solid line) of the position (x6 , y 6 ) of the center of
gravity of link 6 in the (eIx , eIy )-plane during lateral undulation. The model is
simulated with orthotropic friction.
forward mainly along the eIx -axis. The forward motion would
not have been the result for an isotropic friction model (i.e.,
µT x = µT y ). In that case, the snake robot would have moved
slowly backward as we will see in Section VIII-D. In order
to increase the velocity of a snake robot, even with isotropic
friction, the snake robot may push against external objects to
move forward. Such an approach is elaborated on in, e.g., [35]
and [36].
C. Experimental Setup
The snake robot Aiko (Fig. 1) used in the experiments was
built in 2006 at the NTNU/SINTEF Advanced Robotics Labo-
ratory,1 Trondheim, Norway. Aiko has ten 2 DOF cardan joints
and 11 links. The length Li and mass mi of the links are the
same as given for the mathematical model in Section VIII-A.
The metal sphere used as a “face” in Fig. 1 was removed in
the experiments to obtain a uniform weight distribution along
the snake robot. The friction coefficients in the model were
calculated from the measurements obtained by dragging Aiko
from a scale along the particle board. The friction coefficients
were found to be µT x = µT y = 0.2, and these values will be
employed in the simulations in the remainder of this paper.
Each 2 DOF cardan joint was actuated by two 6 W dc motors.
The two motors were controlled by a controller with dynamic
feedback [37] implemented on an Atmel ATmega128 microcon-
troller. The snake robot joint reference angles were sent with a
frequency of 10 Hz from a PC, via a controller area network
(CAN) bus to the microcontrollers. The position of the center
of the middle link (link 6) was tracked using a Vicon MX Mo-
tion Capture System with four cameras (MX3) together with
Matlab Simulink. The Vicon programme (Vicon iQ 2.0) ran on
a computer with four Intel Xeon 3 GHz processors and 2 GB
RAM. The logging of motion data was synchronized in time
through a Transmission Control Protocol (TCP) connection be-
tween the PC that controlled Aiko and the Vicon-computer at
the startup of the transmission of the desired snake robot joint
1 Online Available: http://www.sintef.com/snakerobots
IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 2, APRIL 2008
Fig. 7. Position of link 6 during lateral undulation with isotropic friction on a
flat plane: simulation (dashed line) and experimental (solid line) results.
angles. Data logging was performed at 20 Hz. We have chosen
to let the position of link 6 represent the position of the snake
robot, because then we filter any transient behavior of the snake
robot that might occur at its ends, and it is thus, easier to focus
on the general motion of the snake robot. The center of gravity
of the whole snake robot (instead of a specific link) can also be
advantageous to be used to represent the position of the snake
robot. However, such an approach is cumbersome to realize with
our Vicon system.
Particle boards were used as the ground surface.
D. Lateral Undulation: Simulation and Experimental Results
In this section, we compare simulation and experimental re-
sults for lateral undulation with isotropic friction. The reference
joint angles are calculated from (70) and (71) with the parame-
ters Ah = 40π/180 rad, ωh = 80π/180 rad/s, δh = −50π/180
rad, Av = δv = δ0 = 0 rad, ωv = 0 rad/s, and ψh = ψv = 0,
for both the simulation and the experiment. We let the snake
robot start in a curved posture with its center line approximately
along the eIx -axis and with its initial joint angles given by αh i (0)
and αv i (0). Fig. 7 shows both the position of link 6 calculated
from the mathematical model in the simulation and the position
logged from the experiment with Aiko. We see that the snake
robot moves slowly backward along the eIx -axis. From Fig. 7,
we see that the 3-D model display the same trend in motion as
observed in the experiment. However, the distance traveled is
greater in the simulation, and we have tried varying the friction
coefficients (while still keeping isotropic friction), but this did
not seem to have a noticeable effect on the simulation results.
The difference in distance traveled may come as a consequence
of that the model of the snake robot is able to control its joints
more accurately, and thus, is able to reach the maximum am-
plitude when it is required to do so. Also, this suggests that the
friction contact between the snake robot and the ground surface
is even more complex than what has been accounted for in the
model. A more elaborate list of possible sources of errors is
given in Section VIII-F.
We note from these simulation and experimental results that
the direction of locomotion by lateral undulation with isotropic
TRANSETH et al.: 3-D SNAKE ROBOT MOTION: NONSMOOTH MODELING, SIMULATIONS, AND EXPERIMENTS 373

Fig. 8. Top-down view of snake robot postures during simulation of startup Fig. 9. Aiko (bottom) and simulated snake robot (top) during sidewind-
with the “soft-start” approach for sidewinding. ing. The images show the snake robot at t = 0 s (straight posture), t = 5 s,
t = 10 s, and t = 15 s.

friction is the same as the direction of the waves that is propa-

gated along the body of the snake robot. This is coherent with
results found in [38], where this phenomenon is elaborated on.

E. Sidewinding: Simulation and Experimental Results

In this section, we present and compare the simulation results
and the experimental results for the motion pattern sidewinding.
The joint reference angles are found from (70) and (71) with
the parameters Ah = 30π/180 rad, ωh = 80π/180 rad/s, δh =
−50π/180 rad, Av = 10π/180 rad, ωv = ωh , δv = δh , δ0 =
90π/180 rad, and ψh = ψv = 0 rad, for both the simulation
and the experiment. We employ a “soft start” approach during
the startup of the snake robot from its initial (straight) posture
to avoid large steps in the reference signal. To this end, we
override the expressions (70) and (71) and set αh i (t) = 0 rad Fig. 10. Position of link 6 during sidewinding on a flat plane: simulation
(dashed line) and experimental (solid line) results. The triangle and the circle
until the reference signal is within |αh i (t)| ≤ 3π/180 rad for mark the end of the paths.
the first time after startup. The same applies for αv i . The shape
of the snake robot during startup with the “soft-start” approach
is illustrated in Fig. 8 for sidewinding. model in Matlab in order to keep the time it takes to implement
Fig. 9 illustrates how Aiko and the model move during the 15 s the model relatively short.
simulation of sidewinding locomotion. Since all the simulations Figs. 10 and 11 display the position of link 6 for the 3-D model
have now been performed, a note on the computational cost is and Aiko during sidewinding. We observe from the figures that
appropriate. All simulations were performed on a Pentium-M the model almost follows the eIy -axis, while Aiko steadily turns
1.8 GHz computer running Matlab R2006a. For 20 s simu- somewhat to the right. In addition, the model covers a greater
lations, it took on average about 7.1 min to simulate 1 s of distance than Aiko. We see from Figs. 10 and 11(b) that the sim-
sidewinding motion and only 1.3 min to simulate 1 s of lateral ulation results have the same trend and approximate frequency
undulation. The difference in simulation speed arises from that as the experimental results along the eIy -axis. The variance in
the 3-D motion during sidewinding is more complex than pla- trend is more noticeable along the eIx -axis due to the turning mo-
nar motion by lateral undulation. The simulation speed can be tion of Aiko. Moreover, a slight initial difference between the
greatly increased by instead implementing the model in, for ex- orientation of Aiko and the model may contribute to the discrep-
ample, C++. However, we have chosen to first implement the ancy in heading. We discuss in Section VIII-F various reasons
374
Fig. 11. Position of link 6 during sidewinding on a flat plane:simulation
(dashed line) and experimental (solid line) results.
for the differences between the simulation and experimental re-
sults. For sidewinding, we believe that one of the most important
differences between the 3-D model and Aiko is that Aiko is not
able to control its joints accurately. This is particularly the case
for vertical (lifting) motion that is needed for sidewinding. By
inspecting Aiko closely during the experiment, we noticed that
a large part of its body was touching the ground at the same
time. This was not the case in the simulation where the snake
robot joints were accurately controlled that resulted in that most
if its body was lifted from the ground during locomotion.
We observe from Figs. 10 and 11(b) that Aiko sometimes
slides a little backward (at t ≈ 5 s, t ≈ 9 s, and t ≈ 14 s). We
have tried to reproduce this phenomenon in the 3-D model by
lowering and increasing the friction coefficient. We have also
tried a variety of anisotropic friction properties, but without any
luck. The backward motion might be a result of Aiko not being
able to lift its links properly. The links that are supposed to be
lifted and moved forward are instead sometimes dragged along
the particle board. This results in a friction force that acts in
the opposite direction of the desired motion of the snake robot.
This friction force might result in that some of the parts that
are supposed to be lying stationary on the ground are instead
pushed slightly backward.
F. Discussion of the Experimental Validation
We see from Figs. 10 and 11 that the simulation and exper-
imental results compare fairly well. However, the differences
between the two results need to be addressed, and we list the
most important possible sources of errors in the following. These
effects have not been included in the mathematical model.
1) Aiko has a noticeable free play in the joints of about 3–5◦ .
This results in that the control of the joint angles is not
completely accurate and a joint angle might not be able to
reach its desired angle.
2) The dynamics of the actuators are not modeled. Hence,
the actuators in the models are extremely strong and fast,
and we are able to accurately control the actual angle of
each joint to its desired position. This is not the case for
Aiko where the joint motors sometimes saturate and are
IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 2, APRIL 2008
unable to track the desired joint movement precisely. This
particularly the case for vertical motion where larger joint
torques are required.
3) The exoskeleton of Aiko is not modeled precisely since
this would require a great deal of consideration of the ge-
ometry of the snake robot, and this would severely com-
plicate the expressions for contact with the environment.
4) For each cardan joint, there is, in fact, a 2.2 cm distance
between the rotational axes for the yaw and pitch motion.
However, to simplify the kinematics, we have assumed
that the axes intersect.
5) The Stribeck effect [22] is not included in the model of
friction. The Stribeck effect states, roughly speaking, that
friction forces acting on a body is reduced just after it is
set in motion.
6) The varnish on the underside of the snake robot has been
worn down. This has resulted in that the snake robot links
slide easier when rolled slightly to one of the sides where
the varnish is still present.
7) There was a time delay between the startup of the snake
robot and the startup of the logging of position data. This
delay ranged between 50 and 150 ms.
8) The physical parameters of the snake robot may be slightly
incorrect, partly since we assume that each link has a
uniform mass distribution.
The list of possible errors is long, and it is difficult to de-
termine what issues are the most important in order to explain
the differences between the simulation and experimental results.
Also, the main factors may vary depending on which motion pat-
tern is tested. To this end, we believe that for lateral undulation
1) and 2) are the most important reasons for the differences since
the snake robot is always lying flat on the ground. Moreover, for
sidewinding locomotion, 1)–5) affect the comparison the most.
Even though there are several sources of error, we see from
the plots that compare the simulation and experimental results
that the model gives a satisfactory qualitative description of the
snake robot dynamics. The model is not accurate enough to
precisely predict the quantitative motion of the snake robot, but
it should be possible to improve the accuracy of the model by
taking into account the various sources of errors listed before.
However, at its current state, the model gives a clear indication
of how the snake robot will move during sidewinding and lateral
undulation with isotropic friction and extending the model will
require a considerable amount of work in tailoring the model to
the specific snake robot Aiko. Hence, the process will have to
be repeated for new snake robot designs. Instead of presenting a
model with such a very close resemblance to Aiko, we present
a model for synthesis and testing of new 3-D motion patterns.
Moreover, the comparisons between the simulation and experi-
mental results suggest that our model is valid for this purpose.
IX. CONCLUSION AND FURTHER WORK
In this paper, we present a 3-D nonsmooth mathematical
model that enables synthesis and testing of 3-D snake robot
motion patterns. Experiments show that the models describe in
a realistic manner how a real snake robot will behave during
locomotion.
TRANSETH et al.: 3-D SNAKE ROBOT MOTION: NONSMOOTH MODELING, SIMULATIONS, AND EXPERIMENTS
The model of the snake robot is developed based on the
framework of nonsmooth dynamics and convex analysis. This
framework allows us to easily incorporate the contact forces with
the ground, together with an accurate description of the spatial
Coulomb friction. In addition, even though we employ a hybrid
model, there is no need for an explicit switch between system
equations (when, for example, an impact occurs) since both the
nonimpulsive forces and the impact impulses are covered by
the same force law together with that we use the time-stepping
method for numerical treatment. The use of nonminimal abso-
lute coordinates results in a constant and diagonal mass matrix
and an effective way of writing the system equations. Such a
constant mass matrix is beneficial in the numerical treatment
since it needs only to be inverted once and not in each time step
during simulation.
Simulations of the snake robot during the serpentine motion
pattern lateral undulation is performed. The simulation result
shows that the orthotropic friction model based on Coulomb’s
law of dry friction is reasonable since the snake robot moves
forward.
Experiments are performed with the snake robot Aiko in
Fig. 1 for the serpentine motion patterns lateral undulation and
sidewinding with isotropic friction. Back-to-back comparisons
between simulation results and experimental results with these
motion patterns are given to validate the mathematical model.
The simulation and experimental results compare satisfactorily.
The simulation results together with the experimental valida-
tion show that the mathematical model presented in this paper
gives a satisfactory description of how our snake robot moves in
the real world. The model cannot be employed directly (as, e.g.,
a state estimator) to predict the exact motion of our snake robot.
However, the comparisons with experimental results show that
the model is suitable for developing and testing motion patterns
in order to see how a real snake robot will move for a given
motion pattern.
We show in this paper how to develop a 3-D mathematical
model of a snake robot on a flat ground surface. Further work
will consist of extending the model to include other ground
shapes such as stairs. Also, the model will be used to develop
and test new 3-D motion patterns for snake robots. Moreover,
the optimization of nonsmooth systems is a growing field of
research and the 3-D model presented in this paper may, one
day, be employed to optimize gaits with respect to, for example,
speed or energy efficiency.
It is hoped that this paper can inspire other communities
working on robot manipulators to try out the powerful modeling
techniques available in the framework of nonsmooth dynamics
and convex analysis.
REFERENCES
[1] S. Hirose, Biologically Inspired Robots: Snake-Like Locomotors and
Manipulators. Oxford, U.K.: Oxford Univ. Press, 1993.
[2] S. Hirose and A. Morishima, “Design and control of a mobile robot
with an articulated body,” Int. J. Robot. Res., vol. 9, no. 2, pp. 99–114,
1990.
[3] P. Liljebäck, Ø. Stavdahl, and K. Y. Pettersen, “Modular pneumatic snake
robot: 3D modelling, implementation and control,” presented at the 16th
IFAC World Congr., Prague, Czech Republic, Jul. 2005.
375
[4] M. Mori and S. Hirose, “Three-dimensional serpentine motion and lateral
rolling by active cord mechanism ACM-R3,” in Proc. IEEE/RSJ Int.
Conf. Intell. Robots Syst., 2002, pp. 829–834.
[5] G. Chirikjian and J. Burdick, “Design and experiments with a 30 DOF
robot,” in Proc. IEEE Int. Conf. Robot. Autom., 1993, vol. 3, pp. 113–119.
[6] G. Chirikjian and J. Burdick, “The kinematics of hyper-redundant robot
locomotion,” IEEE Trans. Robot. Autom., vol. 11, no. 6, pp. 781–793,
Dec. 1995.
[7] S. Ma, Y. Ohmameuda, K. Inoue, and B. Li, “Control of a 3-dimensional
snake-like robot,” in Proc. IEEE Int. Conf. Robot. Autom., vol. 2, Taipei,
Taiwan, Sep. 2003, pp. 2067–2072.
[8] J. Burdick, J. Radford, and G. Chirikjian, “A ‘sidewinding’ locomotion
gait for hyper-redundant robots,” in Proc. IEEE Int. Conf. Robot. Autom.,
May, 1993, pp. 101–106.
[9] A. A. Transeth and K. Y. Pettersen, “Developments in snake robot
modeling and locomotion,” in Proc. IEEE Int. Conf. Control, Autom.,
Robot. Vis., Singapore, Dec. 2006, pp. 1393–1400.
[10] S. Ma, N. Tadokoro, B. Li, and K. Inoue, “Analysis of creeping locomo-
tion of a snake robot on a slope,” in Proc. IEEE Int. Conf. Robot. Autom.,
Sep., 2003, pp. 2073–2078.
[11] K. McIsaac and J. Ostrowski, “Motion planning for anguilliform locomo-
tion,” IEEE Trans. Robot. Autom., vol. 19, no. 4, pp. 637–625, Aug. 2003.
[12] M. Saito, M. Fukaya, and T. Iwasaki, “Serpentine locomotion with robotic
snakes,” IEEE Control Syst. Mag., vol. 22, no. 1, pp. 64–81, Feb. 2002.
[13] S. Ma, “Analysis of creeping locomotion of a snake-like robot,” Adv.
Robot., vol. 15, no. 2, pp. 205–224, 2001.
[14] I. Tanev, T. Ray, and A. Buller, “Automated evolutionary design,
robustness, and adaptation of sidewinding locomotion of a simulated
snake-like robot,” IEEE Trans. Robot., vol. 21, no. 4, pp. 632–645, Aug.
2005.
[15] P. Lamon, T. Thueer, R. Jordi, and R. Siegwart, “Modeling and optimiza-
tion of wheeled rovers,” presented at the 8th ESA Workshop Adv. Space
Technol. Robot., New Orleans, LA, 2004.
[16] S. Ma, Y. Ohmameuda, and K. Inoue, “Dynamic analysis of 3-dimensional
snake robots,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2004,
pp. 767–772.
[17] F. Matsuno and H. Sato, “Trajectory tracking control of snake robots
based on dynamic model,” in Proc. IEEE Int. Conf. Robot. Autom., 2005,
pp. 3029–3034.
[18] P. Prautsch, T. Mita, and T. Iwasaki, “Analysis and control of a gait of
snake robot,” Trans. Inst. Electr. Eng. J. Ind. Appl. Soc., vol. 120-D,
pp. 372–381, 2000.
[19] B. Brogliato, Nonsmooth Mechanics, 2nd ed. London, U.K.: Springer-
Verlag, 1999.
[20] R. Smith. (2007, Nov. 15). Open Dynamics Engine v0.5 user guide
[Online]. Available: http://ode.org/ode-latest-userguide.html
[21] Ch. Glocker, Set-Valued Force Laws, Dynamics of Non-Smooth Systems
(Lecture Notes in Applied Mechanics). vol. 1, Berlin, Germany:
Springer-Verlag, 2001.
[22] R. I. Leine and H. Nijmeijer, Dynamics and Bifurcations of Non-Smooth
Mechanical Systems (Lecture Notes in Applied and Computational
Mechanics). vol. 18, Berlin, Germany: Springer-Verlag, 2004.
[23] J. J. Moreau, “Unilateral contact and dry friction in finite freedom
dynamics,” in Non-Smooth Mechanics and Applications (CISM Courses
and Lectures), vol. 302, J. J. Moreau and P. D. Panagiotopoulos, Eds.,
Vienna ed.. Austria: Springer-Verlag, 1988, pp. 1–82.
[24] Ch. Glocker and C. Studer, “Formulation and preparation for numerical
evaluation of linear complementarity systems in dynamics,” Multibody
Syst. Dyn., vol. 13, pp. 447–463, 2005.
[25] A. A. Transeth, R. I. Leine, Christoph Glocker, and K. Y. Pettersen,
“Non-smooth 3D modeling of a snake robot with frictional unilateral
constraints,” in Proc. IEEE Int. Conf. Robot. Biomimetics, Kunming,
China, Dec. 2006, pp. 1181–1188.
[26] O. Egeland and J.T. Gravdahl, Modeling and Simulation for Automatic
Control. Trondheim, Norway: Marine Cybernetics, 2002.
[27] C. Le Saux, R. I. Leine, and Ch. Glocker, “Dynamics of a rolling disk in
the presence of dry friction,” J. Nonlinear Sci., vol. 15, no. 1, pp. 27–61,
2005.
[28] R. T. Rockafellar, Convex Analysis (Princeton Landmarks in Mathemat-
ics). Princeton, NJ: Princeton Univ. Press, 1970.
[29] A. A. Transeth, R. I. Leine, Ch. Glocker, and K. Y. Pettersen, “Non-smooth
3D modeling of a snake robot with external obstacles,” in Proc. IEEE Int.
Conf. Robot. Biomimetics, Kunminga, China, Dec. 2006, pp. 1189–1196.
[30] J. Gray, “The mechanism of locomotion in snakes,” J. Exp. Biol., vol. 23,
no. 2, pp. 101–120, 1946.
376
[31] P. Alart and A. Curnier, “A mixed formulation for frictional contact
problems prone to Newton like solution methods,” Comput. Methods
Appl. Mech. Eng., vol. 92, pp. 353–375, 1991.
[32] J. Ostrowski and J. Burdick, “Gait kinematics for a serpentine robot,” in
Proc. IEEE Int. Conf. Robot. Autom., Apr., 1996, vol. 2, pp. 1294–1299.
[33] P. Wiriyacharoensunthorn and S. Laowattana, “Analysis and design of a
multi-link mobile robot (serpentine),” in Proc. IEEE Int. Conf. Robot.,
Intell. Syst. Signal Process., Dec., 2002, vol. 2, pp. 694–699.
[34] C. Ye, S. Ma, B. Li, and Y. Wang, “Turning and side motion of
snake-like robot,” in Proc. IEEE Int. Conf. Robot. Autom., 2004, vol. 5,
pp. 5075–5080.
[35] Z. Y. Bayraktaroglu, A. Kilicarslan, A. Kuzucu, V. Hugel, and P. Blazevic,
“Design and control of biologically inspired wheel-less snake-like robot,”
in Proc. IEEE/RAS-EMBS Int. Conf. Biomed. Robot. Biomechatron., Feb.,
2006, pp. 1001–1006.
[36] A. A. Transeth, R. I. Leine, C. Glocker, K. Y. Pettersen, and P. Liljebäck,
“Snake robot obstacle aided locomotion: Modeling, simulations, and
experiments,” IEEE Trans. Robot., vol. 24, no. 1, pp. 88–104, Feb. 2008.
[37] O. Egeland, Servo Engineering (in Norwegian). Trondheim, Norway:
Tapir, 1993.
[38] M. Nilsson, “Serpentine locomotion on surfaces with uniform friction,”
in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2004, pp. 1751–1755.
Aksel A. Transeth (S’04–M’08) received the M.Sc.
and Ph.D. degrees from the Department of Engineer-
ing Cybernetics, Norwegian University of Science
and Technology (NTNU), Trondheim, Norway, in
2004 and 2008, respectively.
He is currently a Research Scientist in the Depart-
ment of Applied Cybernetics, SINTEF Information
and Communication Technology (ICT), Trondheim,
a Norwegian research organization. His current re-
search interests include modeling and control of dy-
namical systems, biologically inspired locomotion,
and robotics.
Remco I. Leine received the M.Sc. degree in me-
chanical engineering from the Technical University,
Delft, Delft, The Netherlands, in 1996, and the Ph.D.
degree in mechanical engineering from the Technical
University Eindhoven, Eindhoven, The Netherlands,
in 2000.
He has been a Postdoctoral Researcher with the
Technical University Munich and the Institut Na-
tional de Recherche en Informatique et en Automa-
tique (INRIA) Rhône-Alpes. He has also been a Post-
doctoral Researcher at the Eidgenössische Technis-
che Hochschule (ETH) Zürich, Zurich, Switzerland, where, since 2003, he was
a Senior Lecturer in the Institute of Mechanical Systems, and in 2007, obtained
his Habilitation and was appointed as a “Privatdozent” of mechanics at the
Institut für Mechanische Systeme (IMES) Centre of Mechanics. His current re-
search interests include the intersection of nonsmooth dynamics and nonlinear
dynamics: stability theory for nonsmooth systems, friction-induced vibration
and nonlinear phenomena, discontinuous bifurcations in nonsmooth dynamical
systems, and multibody systems with impact and friction.
IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 2, APRIL 2008
Christoph Glocker received the Diploma and the
Doctoral degree in mechanical engineering, and the
Certificate of Habilitation in mechanics from the
Technische Universität (TU) München, Munchen,
Germany, in 1989, 1995, and 2001, respectively.
In 1996, he was granted a Feodor Lynen Fel-
lowship supported by the Alexander von Humboldt
Foundation and spent one year at Aristotle Uni-
versity of Thessaloniki. In 1997, he joined the TU
München as a Senior Engineer. He then joined the Ei-
dgenössische Technische Hochschule (ETH) Zürich,
Zurich, Switzerland, where he is currently a Full Professor of mechanics with the
Institut für Mechanische Systeme (IMES) Centre of Mechanics. He is engaged
in the research of nonsmooth dynamics of mechanical systems with finite de-
grees of freedom that includes, for example, the friction and impact problem in
multibody systems, providing a generalization of classical analytical mechan-
ics. His current research interests also include both theoretical and practical
questions concerning the mechanical modeling, the mathematical formulation,
and the numerical treatment of systems with discontinuities, as well as their
applications to industrial problems.
Kristin Y. Pettersen (S’93–M’98–SM’04) received
the M.Sc. and Ph.D. degrees in electrical engineering
from the Norwegian University of Science and Tech-
nology (NTNU), Trondheim, Norway, in 1992 and
1996, respectively.
She is currently with the Department of Engi-
neering Cybernetics, NTNU, where she became an
Associate Professor in 1996 and a Professor in 2002.
During 1999, she was a Visiting Fellow in the De-
partment of Mechanical and Aerospace Engineering,
Princeton University, Princeton, NJ. She is the author
or coauthor of more than 80 published conference and journal papers. Her cur-
rent research interests include nonlinear control of mechanical systems with
applications to robotics, satellites, autonomous underwater vehicles (AUVs),
and ships.
Prof. Pettersen received the IEEE Transactions on Control Systems Technol-
ogy Outstanding Paper Award in 2006. She also holds several board positions.